On Wed, Jun 29, 2022 at 02:45:11PM +0200, Björn Töpel wrote:

On Wed, 29 Jun 2022 at 12:58, Maciej Fijalkowski
[off-list ref] wrote:

quoted

When application runs in zero copy busy poll mode and does not receive a
single packet but only sends them, it is currently impossible to get
into napi_busy_loop() as napi_id is only marked on Rx side in
xsk_rcv_check(). In there, napi_id is being taken from xdp_rxq_info
carried by xdp_buff. From Tx perspective, we do not have access to it.
What we have handy is the xsk pool.

The fact that the napi_id is not set unless set from the ingress side
is actually "by design". It's CONFIG_NET_RX_BUSY_POLL after all. I
followed the semantics of the regular busy-polling sockets. So, I
wouldn't say it's a fix! The busy-polling in sendmsg is really just
about "driving the RX busy-polling from another socket syscall".

I just felt that busy polling for txonly apps was broken, hence the
'fixing' flavour. I can send it just as improvement to bpf-next.

That being said, I definitely see that this is useful for AF_XDP
sockets, but keep in mind that it sort of changes the behavior from
regular sockets. And we'll get different behavior for
copy-mode/zero-copy mode.

TL;DR, I think it's a good addition. One small nit below:

quoted

+                       __sk_mark_napi_id_once(sk, xs->pool->heads[0].xdp.rxq->napi_id);

Please hide this hideous pointer chasing in something neater:
xsk_pool_get_napi_id() or something.

Would it make sense to introduce napi_id to xsk_buff_pool then?
xp_set_rxq_info() could be setting it. We are sure that napi_id is the
same for whole pool (each xdp_buff_xsk's rxq info).

Björn